Disconfirming Evidence: Actively Looking for What Breaks the Thesis

Disconfirming evidence is the observation named in advance that would show a view is wrong. For the worked claim below it is gross margin at 46.0 per cent, held across four consecutive quarters, with input cost per unit rising through them, or the 0.13 percentage point share gain repeating. Named after the view is held, it stops being a test and becomes a defence.

A breaking observation rests on three ideas already settled elsewhere. The material on the three statements settled what a gross marginWhat is left of revenue once materials have been paid for, shown as a percentage of revenue. is and how it moves. The valuation work settled the arithmetic that runs backwards from a quoted price to the assumptions buried inside it. Thesis variables settled that the worked claim about Sarvani Coatings Limited rests on only three variables. Catalysts settled that a catalystA dated event that could prompt other readers to revise their own assumptions. The event is about the audience rather than about the business. is an event that might move other people rather than evidence about the business. Disconfirming evidence is the other half of a view: not what it rests on, but what would knock it over, written down at a moment when nobody yet knows whether it will ever happen.

What is disconfirming evidence, and how is it different from a risk?

The two words come apart cleanly. Evidence, in the ordinary sense: something that can be looked at. Disconfirming: it counts against the claim rather than for it. Put together, disconfirming evidence is a specific thing which, if it were seen, would mean the claim was mistaken.

A breaking observation is not a risk, and the difference matters more than it sounds. A risk is a thing that could go badly. A breaking observation is a thing that, if it happened, would prove the claim wrong. The two sentences have different grammar, and only one of them can be settled by looking. Nothing on a list of risks is an observation the claim itself said would not occur, so a list of risks leaves a view completely untested.

Here is the household version, and it is worth sitting with before the finance one arrives. A parent says the new tuition centre is working, and lists what could go wrong: the teacher might leave, the batch might get too big, the commute might become a problem. Every item is real and every item is worth watching. None of them is the parent saying in advance what a result would have to look like for the claim to be wrong. Not one of them is a test. The test would be a single sentence written in June: if the December marks in mathematics are not above the June ones, the centre is not working. Now the claim can lose. Before that sentence existed, it could not.

Why does it have to be written before the view is held?

The honest version of the question comes first: what would actually happen if the breaking observation were written six months in, after living with the view through two results? Nothing dramatic. The analyst would sit down, think carefully, and write something reasonable. And it would be something the view could survive.

Not because anybody cheated. Six months brings a working knowledge of which numbers wobble and which hold steady, and every candidate observation arrives already carrying a rough sense of how likely it is to fire. The ones that feel dangerous feel, quite genuinely, like the wrong tests. There is always a reason a dangerous test is the wrong test, and the reason is usually a good one. Evidence picked once the view is already in place turns out, every time, to be evidence the view can live through. No dishonesty is needed and nobody decides on it. The remedy is therefore a rule about order rather than an appeal to being careful.

The everyday shape of this is a household setting a budget. In April nobody knows which month will be expensive, so a household that decides then that the second scooter goes if petrol for it ever passes Rs 2,000/- in a month has a rule with teeth. A household that decides the same thing in October, holding the October bill, is not deciding a rule. The October household is deciding whether Rs 2,240/- is a lot, with a specific and very reasonable explanation for that particular month already in front of it. The second household is not weaker than the first. The order of the work is simply wrong.

What actually counts as disconfirming, and what only looks like it?

Here is the test, and it is a single sentence. The observation must be one the claim itself said would not happen, and anything else is merely disappointing.

The second half of that sentence is where readers go wrong, and they go wrong in the generous direction rather than the lazy one. Suppose the claim is about margin, and a quarter arrives with soft volume. Soft volume is bad news for anybody holding the view. The claim never said volume would be strong in any particular quarter, so soft volume is not evidence against the view. Counting it as disconfirming feels rigorous and is the opposite: it widens the test until almost any unpleasant fact qualifies, and a test that fires on almost anything is a test that settles nothing.

Notice what the generous reader has actually done. The reader has handed themselves a way to retire the claim on evidence the claim never bet against, and can say afterwards that the view was tested. The claim never lost. The claim was withdrawn.

So a counterfeit is not usually a lie. A counterfeit is a real fact, correctly measured, that simply sits outside the sentence that was written. A useful habit is to read the arriving fact and the original claim side by side, out loud, and ask whether the second one forbade the first. If it did not, the fact is information about the company and it is not information about the claim.

What does the claim under test actually say?

Everything from here uses one worked claim about Sarvani Coatings Limited, an invented listed maker of decorative paints and industrial coatings. The claim is written out so that a reader has something specific to pull apart. A claim of this kind is not a target, a rating or a statement of what a share is worth. It is one sentence about one mechanism, built so that it can be shown wrong.

Two observations were named against that claim, and both were written into the record before either could be seen. Writing them before either could be seen is not decoration. Writing them first is the entire reason they are worth studying.

Why does the first observation need every clause it has?

The sentence is short and every part of it is load bearing. Take the clauses one at a time and watch what happens when each is removed. Nothing shows what a well built observation is doing faster.

Start with four consecutive quarters. Sarvani Coatings' year has a strong seasonalityThe pattern that makes one quarter genuinely incomparable with the quarter immediately before it. The two quarters carry different trading conditions. Set up in the earnings material., and the published quarters show it plainly. In year three the quarterly earnings before interest, tax, depreciation and amortisation (EBITDA) margins were 17.97, 16.15, 20.71 and 18.45 per cent against a blended marginThe figure obtained by taking the whole year's profit over the whole year's revenue, rather than averaging the four quarterly percentages. The two are not the same number. of 18.47 per cent for the full year. The four figures spread 4.57 points inside one ordinary year. The third quarter carries the festive season and the second carries the monsoon, so a single quarter can move several points for reasons that have nothing whatever to do with whether the company can price. Remove the four quarter clause and the test fires on the season instead of on the claim.

Now the second clause, with input cost per unit rising. The direction of the input is the harder condition, and it is what makes the observation work at all. Across a single year, from year two into year three, gross margin rose 2.0 points from 44.0 to 46.0 per cent while materials fell from 56.0 to 54.0 per cent of revenue. On volume up 6.0 per cent, realisationRevenue divided by the units actually sold, so the average price a maker got per unit. Built and used in the earnings material rather than here. per unit gained about 7.5 per cent while materials cost per unitWhat one unit of output cost to make, being the cost of materials divided by the units produced. Materials as a share of revenue also moves when price moves, so the two figures are different. gained about 3.6 per cent. Multiplying 56.0 per cent by 1.0363847 over 1.0746707 gives 53.9959. The published figure is 54.0. So the margin gain came from realisation outrunning the input. The input itself did not get cheaper.

Now take that condition out and see what happens to the test. Suppose the margin stays at 46.0 per cent through four quarters while the input gets cheaper. Has the claim been broken? No. The claim has been confirmed. A margin held on a cheapening input is exactly what a single step up looks like: the company is enjoying a cheaper input and passing none of it back, and nothing has been shown about whether it can price when the wind turns. Dropping that condition does not merely loosen the test, it inverts it, turning an observation that would break the claim into one the claim comfortably predicts.

The inversion is worth naming plainly. Removal only ever loosens a test, so no removal can ever make one impossible to pass. Drop the period and the test fires on the season. Drop the direction of the input and the test turns around to point the other way. Two different failures, and the second is much the more instructive.

Why is a 0.13 percentage point share gain worth naming as the second test?

Because it is small, and because its smallness is part of the observation rather than an embarrassment to be smoothed over. The arithmetic is worth working rather than accepting the sentence. Sarvani Coatings' Rs 2,415 crore of revenue divided by an invented field of Rs 48,300 crore gives 5.00 per cent. The same division for the year before, Rs 2,120 crore over Rs 43,500 crore, gives 4.8736. The figure prints as 4.87. The gain is 0.13 percentage pointsThe straight subtraction of one percentage from another. A share moving 4.87 up to 5.00 means 0.13 percentage points. Saying it rose 2.6 per cent is a different statement..

Now state the identical event the other way. Sarvani Coatings put on 13.9 per cent of revenue growth against a field that grew 11.0 per cent, so it outgrew the field by 2.9 percentage points. Both sentences are true, both describe the same year, and the second number is 22.8 times the first. Both statements are accurate and only the share figure keeps the size of the event honest. The observation is therefore written in share and not in outgrowth.

There is a second reason two observations beat one, and it is not that two is safer. A margin holding and a share gain repeating are different kinds of evidence about different mechanisms. The first is about whether the company can price. The second is about whether it is winning volume against everybody else. A single piece of evidence can fail to arrive for boring reasons: a disclosure changes, a definition moves, a year is odd. Two observations watching two mechanisms mean the claim can lose in two independent ways. Two independent ways is a stronger commitment than one test twice as strict.

What does a counterfeit look like beside the real ones?

Put a third sentence next to the two named observations: a quarter of soft volume. Soft volume is a real fact when it arrives. The number is measured the same way as everything else. The quarter would make anybody holding the claim uncomfortable. And it contradicts nothing whatsoever in the claim. The claim is about margin being a step rather than a rate, and says nothing at all about how many litres get sold in any three month stretch.

The counterfeit offers an exit. Treat it as disconfirming and the holder gets to retire the claim on evidence the claim never bet against. The retirement sounds rigorous. The claim was never tested. The claim was quietly dropped and given a respectable reason.

What happens on the day the observation actually arrives?

There is an order, and the order is short enough to memorise. First, the analyst confirms the observation is the one that was named, in its own words, clause by clause. Second, the analyst checks the period. Half the arguments about whether a test fired are really arguments about which quarters were being counted. Third, the analyst writes down that the claim was wrong, beside the claim itself.

The third step is the one people skip, and skipping it is worse than never writing the observation at all. A claim quietly amended leaves the record holding a view that nobody ever actually held, so it is worse than a claim abandoned. Reopened in a year, it holds a sensible, moderate position with no history, and no way to tell whether it was reached before the evidence or assembled afterwards out of what survived. The point of keeping a file is to learn from it, and a file like that teaches its writer nothing.

How does a view defend itself when nobody has decided to defend it?

Three things happen, none of them deliberate, and they compound. New facts get read in the light of the claim, so a neutral number arrives already tilted. The strongest contrary point picks up a reason on the way in, and the reason is usually correct. Being correct is what makes the reason invisible. And the test itself widens, one small reasonable step at a time, until the sentence being checked against bears almost no relation to the one that was written.

Watch the widening happen in slow motion. The observation starts as 46.0 per cent, four quarters running, input getting dearer. A quarter comes in at 45.6. A tenth or two is surely noise, so the working phrase becomes 46.0 give or take a little. Another quarter and it is broadly holding up. By the fourth it is no real deterioration. No year on record has ever failed that test. Nobody made a decision at any point. Every single step was defensible on its own.

A sentence fixed in the record cannot be widened after the fact without the widening being visible, so a written observation is the only mechanical defence against the drift. Willpower is not a defence, and neither is being unusually honest. The analyst doing all of the above is careful, and being told to try harder gives that person nothing to do on Monday morning. Comparing this quarter's working phrase against the words in the file is something they can actually do.

How specific does a breaking observation have to be?

Specific enough that somebody who disagrees with the claim could settle it. The whole standard comes down to three items: a quantity, a period and a direction. A gross margin of 46.0 per cent gives the quantity. Four quarters running gives the period. An input that keeps getting dearer gives the direction, and it is the direction that decides which way the observation cuts.

The check is a handover. The sentence goes to a colleague who thinks the claim is wrong, along with a single question: could this be settled, on a date, either way? If the colleague hesitates, it is not because they are being difficult. The hesitation means there is nothing in the sentence for them to settle, so the observation has not been written yet, however carefully it has been thought about.